POTENTIAL AND LIMITS OF AIRBORNE REMOTE SENSING DATA FOR
EXTRACTION OF FRACTIONAL CANOPY COVER AND FOREST STANDS AND
DETECTION OF TREE SPECIES
L. T. Waser a, H. Eisenbeiss b, M. Kuechler a, E. Baltsavias b
a
Swiss Federal Research Institute WSL, Department of Landscape Inventories, 8903 Birmensdorf,
Switzerland+41 44 739 22 92-(waser, kuechler)@wsl.ch
b
Institute of Geodesy and Photogrammetry, ETH Zurich, 8093 Zurich, Switzerland+41 44 633 30 42(ehenri, manos)@geod.baug.ethz.ch
Commission VIII, ThS-21
KEY WORDS: Forestry, Ecosystem, LIDAR, Modelling, DEM/DTM, Aerial, High resolution, Multisensor
ABSTRACT:
This study presents a methodology for derivation of fractional canopy cover, detection of main tree species, and extraction of forest
stands using logistic regression, airborne remote sensing data and field samples. In a first step, canopy height models (CHMs) are
generated using medium point density LiDAR DSM and DTM and a high-quality matching DSM. Then, fractional canopy covers
are calculated using logistic regression models and explanatory variables from LiDAR and matching CHM, whereas the latter
produced better results due to higher quality and was therefore further used in this study. Based on this fractional canopy cover,
main tree species and forest stands are modelled using logistic regression and airborne digital sensor data ADS40 and CIR aerial
image data as input variables. Good accuracy for the extraction of canopy cover, distinction between coniferous and deciduous trees
and classification of five main tree species (kappa = 0.7 to 0.9) were obtained but classification of additional three deciduous tree
species was less accurate. The extraction of forest stands produced visually satisfactory results but this method suffers from some
limitations and further research is needed. The present study reveals that the extracted forest attributes may be helpful to support
stereo-image interpretation and field surveys in the frame of the Swiss National Forest Inventory (NFI) and may also be useful for
updating existing forest masks and forest management and protection tasks.
1. INTRODUCTION
The present study was carried out in the framework of the Swiss
Mire Monitoring Program and the Swiss National Forest
Inventory (NFI). Since tree growth and expansion of forest
areas impact the non-forest areas of a protected mire biotope,
especially the extraction and modelling of the above mentioned
forest attributes is essential (Küchler et al., 2004, Waser et al.,
2008).
Recent progress in three-dimensional remote sensing mainly
includes digital stereo-photogrammetry, radar interferometry
and LiDAR (Watt and Donoghue, 2005, Baltsavias et al., 2007).
E.g. by subtracting a DTM from the corresponding DSM,
canopy height models (CHMs) can be calculated that serve as
basis for other forest attributes. Using digital photogrammetry,
DSMs are generated via image matching, often using crosscorrelation (Hyyppä et al., 2000) or less frequently multi
image-matching approaches (Zhang and Gruen, 2004).
Meanwhile several LiDAR systems are commercially available
(Naesset and Gobakken, 2005), enabling the derivation of
DTMs from such data as well (Baltsavias, 1999). Several
studies have integrated LiDAR with optical remotely sensed
data to estimate forest attributes such as stand composition, tree
height, crown diameter, basal area, and stem volume (e.g.
Straub, 2003; St-Onge et al., 2004, Baltsavias et al., 2008).
Combining some of these attributes can be useful to evaluate
extent of forest area, detect changes in forest stands (Waser et
al., 2008b), and determine tree/shrub species (Holmgren and
Persson, 2004). According to Scott et al. (2002) modern
regression approaches have proven particularly useful for
modelling spatial distribution of tree species and communities.
Thus, high-resolution remote sensing data in combination with
regression analyses are promising for modeling forest
composition and tree species (e.g. Lamonaca et al., 2008). The
objective of this study is to develop a methodology for
derivation of fractional canopy cover, detection of main tree
species and derivation of forest stands using logistic regression,
airborne remote sensing data and field samples.
2. MATERIAL AND METHODS
2.1 Test site
Models have been developed and tested for a representative
forest ecosystem in the northern Pre-alpine zone of Switzerland.
The test site “Breitmoos” (approx. 47°18’ N and 9°14’ E) has
an extent of approx. 2 km2 and is characterized by a varying
terrain, mixed land cover and deciduous and coniferous forests
with a core mire area (see Fig. 1).
2.2 Remotely sensed data
In this study, four different data-sets are available:
- Leica RC30 frame camera, 4 CIR (red, green, near-infrared
bands) aerial images (1 strip) of 2005, scale 1:5,600 and an
orthoimage that was generated with a spatial resolution of
0.25 m. Main disadvantage: The images are only available on
request and do not cover entire Switzerland.
1405
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
- A DSM that was generated automatically from the above
images with a spatial resolution of 0.25 m.
- National raw LiDAR DSM and DTM data from 2003, leaveson .
- Leica airborne digital sensor (ADS40) images Level 1
(rectified) of 2005, scale 1:30,000, RGB bands (16 bit), spatial
resolution 0.25m. An orthoimage was generated using the
LiDAR DSM. Main advantage: The image data is available
for the area of entire Switzerland every three years.
cloud co-registration procedure as described in Akca (2005) and
Gruen and Akca (2005). This co-registration uses a 7-parameter
3D similarity transformation to remove systematic differences
(bias) between the datasets, e.g. due to different image
orientation. For the estimation of these parameters, control
surfaces were used, i.e. DSM parts that did not change in the
two datasets, i.e. bare ground, and large differences due to
matching errors were removed with a robust filtering. After coregistration of the 2003 and 2005 DSMs, the Z component of
the Euclidian distances (sigma a posteriori) was 2.8 m, showing
a clear reduction of trees and other wooded plants from 2003 to
2005 (see Fig. 3 in the result section for details).
Figure 1. Test site “Breitmoos” with typical mixed forests
(Pixelmap © 2006 Swisstopo JD052552)
Orientation of the CIR aerial images was performed in LPS and
the following accuracies were obtained. The aposteriori sigma 0
was 0.32 pixel and the RMSE for the control points: X=0.06 m,
Y=0.06 m, Z=0.17 m and RMSE for the check points are
X=0.06 m, Y=0.06 m, Z=0.24 m. High-resolution DSM data is
indispensable since accurate surface information of the forest
area is very important for modelling forest composition, e.g.
tree species. Thus, a matching method was applied that can
simultaneously use any number of images (> 2). It is
implemented
in
the
operational,
quasi-complete
photogrammetric processing package Sat-PP which supports
satellite and aerial sensors with frame and linear array geometry
(for further details see Zhang (2005) and Zhang and Gruen
(2004)). In principle, the matching method consists of three
mutually connected components: an image pre-processing, a
multiple primitive multi-image matching (MPM) and a refined
matching procedure. This automated DSM generation provides
high accuracy (1-5 times the ground sampling distance (GSD),
depending on landcover and terrain roughness), and enables to
produce very dense (grid spacing = 3-4 x GSD) and detailed
DSMs that allow a good 3D modeling of trees and shrubs (see
Fig. 2).
From the raw LiDAR data (first and last pulse), both a DTM
and DSM were generated by the Swiss Federal Office of
Topography (SWISSTOPO). The average density of the DSM
data was 1-2 points / m2 and the height accuracy (1 sigma) 0.5
m for open areas and 1.5 m for vegetation. The DTM has an
average point density of 0.8 points / m2 and height accuracy (1
sigma) of 0.5 m (Artuso et al., 2003). The raw LiDAR data
provided by SWISSTOPO were converted using SCOP++ V5.3
(INPHO) to a raster with 1 m resolution. For the interpolation
of the LiDAR DTM the classic prediction method with default
parameters, while for the LiDAR DSM the robust filtering with
the so-called “LiDAR DSM parameter” was used. For
comparison of the different input data, the matching DSM and
the LiDAR DSM and DTM were co-registered using a point
Figure 2. DSM of the entire test site generated by the SAT-PP
matcher. The DSM is coded in color from light green (light In
BW) to cyan (dark in BW) for lowest and highest elevations
respectively.
2.3 Training and reference data-sets
A ground survey was carried out in summer 2006 and 2007
within the test site. Since the LiDAR data was acquired in 2003
and the image data in 2005 only trees which are visible in both
data-sets were considered as training and reference objects. In
total, for 480 trees we collected: i) tree positions using a subdecimeter GPS with differential correction, ii) tree heights using
a tachymeter and iii) determination of eight tree species. These
are: alder, maple, birch, beech, ash, sorbus (all deciduous trees)
and white fir, spruce (coniferous trees). This information was
used for calibration and validation of the models.
2.4 Fractional canopy covers
The forest / non-forest decision is based on several steps: In a
first step, two canopy height models (CHM) were produced
(CHM1 = matching DSM – LIDAR DTM; CHM2 = LiDAR
DSM –LiDAR DTM). Then both CHMs were used to extract
potential tree areas according to the > 3 m height definition of
trees in the Swiss national Forest Inventory (NFI) (Brassel and
Lischke, 2001). In a second step, non-tree objects (buildings,
rocks etc.) of both CHMs were removed using normalized
difference vegetation index (NDVI) information (low values)
obtained from the CIR aerial images. In a third step, based on
the canopy covers two fractional shrub/tree covers were
1406
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
produced using a logistic regression approach (e.g. McCullagh
and Nelder, 1983) with a probability for each pixel to belong to
the class “forest”. The explanatory variables consist of five
commonly used topographic parameters derived from the
CHMs (slope, aspect, curvature, and two local neighbouring
functions). This approach and the extraction of explanatory
variables are described in detail in Waser et al. (2007) and
Waser et al. (2008a). Using the fractional shrub/tree covers,
various cover strata were computed using different probability
ranges (see Table 1).
and the dominating tree heights per segment (tree group or tree
object) were calculated using moving window approaches.
Since no full-waveform LiDAR data was available for this test
site the third parameter multi-layer information (structure)
could not be achieved.
2.5 Forest composition
The predicted shrub/tree cover strata were validated using a
pixel-to-pixel comparison on the 240 randomly sampled
reference field measurements. For this validation the
corresponding pixel clusters (5x5 pixel window) of the
measured 240 tree/shrub samples and 150 non-tree samples,
respectively were used. Table 1 presents the correspondence
between the pixels of randomly sampled shrubs/trees that are >
3 m and the modelled individual shrub/tree cover strata. The
following statistical measures were used: correct classification
rate (CCR), consumer’s accuracy, producer’s accuracy, kappa
coefficient and correlation coefficient (r2). The accuracies for
five different tree/shrub cover strata are given for both CHMs.
Table 1 clearly shows that generally better accuracies are
obtained by using the topographic parameters from the
matching DSM as explanatory variables. The most accurate
cover stratum also varies. Best correspondence between the
models and field data is obtained for cover stratum 20-100%
when using the matching DSM variables, and 30-100% when
using the LiDAR DSM variables.
Forest composition is commonly regarded as a combination of
degree of composition (i.e. separation on deciduous and
coniferous trees) and tree species.
Before modelling the degree of composition, the tree species
and forest stand results were smoothed using ArcGIS and
segmentation of the ADS40 images was applied using the
Definiens developer 7.0 software. The segmentation provides
groups of trees and single trees with similar shapes and spectral
properties. Distinction of coniferous / deciduous trees and
distinction of main tree species is primarily based on the
fractional canopy cover – all pixels with a forest probability of
less than 0.2 were skipped.
2.5.1
Distinction of coniferous / deciduous trees
For each pixel within a segmented tree group or tree object,
probability belonging to deciduous or coniferous trees was
calculated using a logistic regression model. As explanatory
variables we used 18 parameters derived from the ADS40
images and the RC30 images: original bands of RGB and CIR,
the ratio of each band divided by the sum of the corresponding
3 bands, and the same ratios smoothed with a 3x3 local average
filter. The last 12 parameters have shown a good performance
based on empirical tests. 240 tree species (half of the field data)
were used as training data.
2.5.2
3. RESULTS
3.1 Canopy covers
Cover
stratum
CCR
Cons. Ac.
Prod. Ac.
Distinction of main tree species
Kappa
Probability belonging to one of the eight main tree species was
calculated using a second logistic regression model for each
pixel within a segmented tree group or tree object. For this
model we used 20 spectral variables, which consist of the 18
parameters used in 2.5.1 plus NDVI and smoothed ADS40 blue
channel. Again, 240 tree species (half of the field data) were
used as training data. As output, probabilities for each tree
species within an image segment were obtained.
2.5.3
Forest stands
According to the Swiss NFI a stand can be defined by the three
parameters composition, diameter at breast height (dbh) and
structure (multi-layers of tree cover). Composition can be
determined using the distinction of coniferous / deciduous trees
and classification of the eight tree species as performed in 2.5.1
and 2.5.2.
r2
10100%
0.841
0.923
0.801
0.817
0.842
0.924
0.721
0.802
0.765
0.832
20100%
0.883
0.971
0.861
0.965
0.863
0.934
0.786
0.915
0.782
0.924
30100%
0.902
0.958
0.902
0.954
0.887
0.912
0.823
0.903
0.821
0.908
40100%
0.891
0.935
0.865
0.956
0.857
0.881
0.782
0.883
0.812
0.891
50100%
0.863
0.921
0.836
0.938
0.835
0.834
0.761
0.856
0.747
0.864
Table 1. Accuracies of both canopy covers for five different
strata. 1st lines are based on the Lidar DSM variables, whereas
second lines were obtained from the matching DSM variables.
Fig. 3 clearly shows the different extent of canopy covers with
tree height classes when a) using the LiDAR CHM and b) using
the matching CHM. For comparison purposes the DSMs were
resampled to 1m. The varying extent of forest area is due to
quality differences of the LiDAR data and the matching DSM
but also due to deforestation between 2003 and 2005.
Since it is not possible to obtain the diameter at breast height
from the input data-sets tree height information and minimum
tree area (500m2) was used instead. Tree height is indirectly
linked with the diameter of a trunk. Four tree height classes
according to the definition by the NFI were built using the
CHMs: 3-8m, 8-15m, 15-25m and > 25m. Minimum tree area
1407
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
3.2 Coniferous / deciduous trees
An example of extracted deciduous and coniferous trees is
given for a typical area within the test site in Fig. 5. The
corresponding accuracies are given in Table 2.
Figure 3a. Extracted forest area and tree height classes based on
the LiDAR CHM of 2003.
Figure 5. Left: ADS40 RGB image, right: the corresponding
distinction of deciduous / coniferous trees using ADS40 input
data.
Lowest accuracies are obtained when using CIR images only.
Best results were obtained using both ADS40 images and CIR
images as input data-sets. Table 2 also reveals an increasing
percentage of coniferous trees (and a decrease of deciduous
trees) when both image data-sets were used. Since this increase
compared to the CIR results is about 10% it might be of interest
for further analysis. However, the differences between using
both datasets and ADS40 only is small.
Figure 3b. Extracted forest area and tree height classes based on
the CHM of the matching DSM of 2005. New deforestation is
visible in the left part of the area.
Since the extracted vegetation obtained by the LiDAR DSM is
less accurate and detailed as by the matching DSM data, for
further investigation only the latter was used. Fig. 4 gives an
example of the fractional canopy cover for a typical area within
the test site using the matching DSM.
%
coniferous
%
deciduous
Based on
CCR
K
CIR
88.1
0.78
57.1
42.9
ADS40
93.8
0.87
63.7
36.3
CIR +
ADS40
94.8
0.90
64.3
35.7
Table 2. Accuracies (CCR and Kappa) of distinction between
coniferous / deciduous trees.
3.3 Tree species
In general, best distinction is obtained when combining both
ADS40 and CIR data as explanatory variables. Low
classification rate (CCR) is obtained when considering all eight
tree species. Therefore the focus was laid on five main tree
species (spruce, white fir, ash, beech, birch). Table 3 reveals the
accuracies using either CIR images, ADS40 images or a
combination of both data-sets and shows that best results are
obtained from the latter. Table 3 also shows that this
improvement is less pronounced for deciduous trees.
based
on
CIR
ADS40
CIR+
ADS40
Figure 4. Calculated tree-/shrub probability as provided by the
fractional canopy cover based on the matching DSM data. The
range 20-85% is not shown as it covered a very small area.
spru
ce
0.75
0.86
0.92
w. fir
ash
beech
birch
K
0.64
0.85
0.88
0.81
0.84
0.91
0.80
0.82
0.86
0.76
0.78
0.84
0.68
0.75
0.86
Table 3. CCR values of the five main tree species as obtained
by logistic regression. Kappa (K) values are the average for all
5 species.
1408
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
3.4 Forest stands
Fig. 6 shows that forest stand classes are a combination of tree
heights and forest composition. However, since no forest stand
maps exist from this region, the quality control of the different
forest stands was performed visually using stereo-image
interpretation. In general, the typical stand classes are well
represented when accepting some discrepancies especially in
stand
classes
with
lower
tree
heights
(3-8m).
Figure 6a. Tree height classes according to the Swiss national
forest inventory (NFI).
Figure 6b. Forest stand classes as obtained by combining tree heights and dominant tree type.
4. DISCUSSION AND FINAL REMARKS
This study highlights the potential of combining airborne
remote sensing data with logistic regression models to obtained
forest attributes such as area, tree species and stands on subpixel level. The usage of standard explanatory variables derived
from the DSMs, as already applied in other studies (Küchler et
al., 2004, Waser et al. 2008a), proved to be a good approach for
modeling fractional canopy covers. Accurate extraction of these
forest attributes might be essential for tasks of the NFI (e.g.
support for stereo-interpretation of aerial images), for forest
owners and also forest management tasks.
The first objective, the extraction of forest area, was achieved
semi-automatically with high accuracy using either LiDAR or
matching DSM data. Generally, the usage of explanatory
variables from the matching DSM produced better accuracies –
especially regarding single tree detection. Highest
correspondence between fractional canopy cover and field data
was obtained for the fractional canopy cover strata of 20-100%
using the matching DSM and 30-100% using the LiDAR DSM,
respectively. However, this validation is based on trees > 3 m
(according to the minimum tree height as defined by the NFI).
It has to be kept in mind that accuracies for the strata may vary
when considering also shrubs/trees < 3 m for validation. But
detailed visual stereo-image interpretation confirmed that most
small trees are then within the forest area when using the two
canopy cover strata. Since this study clearly reveals that
accuracy of fractional canopy covers strongly depends on the
accuracy of the DSM data, newly developed, high-quality
matching methods are indispensable. The usage of a dense and
accurate DSM is an absolute prerequisite in order to be able to
derive accurate topographic parameters which in turn are used
to derive the fractional canopy covers. The LiDAR DSM as
applied in this study has a lower point density than the
matching DSM, and due also to partial canopy penetration
(especially with leaves off) is less accurate for modeling
vegetation canopy. In the future, we will be able to produce
high-quality DSMs also from ADS40 data.
The second objective, the distinction of main tree species, was
achieved semi-automatically and mediocre to high accuracies
1409
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
were obtained. First, simple distinction between deciduous trees
and coniferous trees revealed higher accuracies than the
distinction of several tree species. Second, better results were
obtained when using ADS40 data, especially for coniferous tree
species. Anyhow, distinction of tree species based on CIR aerial
image information at least produced satisfactory results for
spruce, ash, beech and birch and partly satisfactory for white fir.
Best results were obtained when combining both sensors as
input for the logistic regression models. However, both sensors
failed to distinguish further deciduous trees with good quality.
Possible reasons are that alder, maple and sorbus are often
grouped, have smaller crowns (and are therefore partly covered
from each other or from other more dominant species) and also
have very similar spectral properties.
This study clearly reveals that spectral information of the blue
band (428-492 nm) as provided by ADS40 is essential to
distinguish between several deciduous tree species. Another
advantage of the ADS40 sensor is that the three spectral bands
are not overlapping. Two final remarks are given here: i)
advantages of this approach are that only few training data is
needed and this method can be applied semi-automatically –
also for larger regions; ii) an important disadvantage is the
availability of both image data, whereas the CIR aerial images
are only available on request. Unfortunately, with this first
generation ADS40 sensor, the NIR line CCD is placed far away
from the RGB CCDs and thus the NIR image looks very
different and cannot be combined with the RGB images
Therefore, in near future the focus will be laid on single usage
of the 2nd generation ADS40 sensor where all four spectral
CCDs are very close to each other. This will improve the results,
by using also the NIR information. Another advantage will be
the availability of newest ADS40 imagery for entire
Switzerland every three years. In this case, the dependence on
image on request will cease to exist. To make this approach
more valuable and applicable for e.g. entire Switzerland it is
also planned to test other tree species of different geographical
regions (e.g. Central and Southern Alps). Another approach will
be the usage of tree texture properties and of the seasonal
variability of tree species using multi-temporal data.
The third objective, the extraction of forest stands, produced
visually satisfactory results but still suffers from some
limitations. Structure, the third necessary parameter to define a
stand, had to be ignored since it can only be derived from fullwaveform LiDAR data. Another drawback is that we used tree
height classes supposing a relation to the diameter at breast
height. Nevertheless, we are able to extract semi-automatically
stand classes that will help for a better understanding of the
structure of forest.
To summarize, this study clearly shows the potential and the
limits of the three extracted forest attributes: area, composition
and stand using high-resolution airborne remote sensing data.
These forest attributes may be helpful to support some tasks in
the Swiss NFI (e.g. stereo-image interpretation, field surveys).
Furthermore, this information could also be useful for updating
existing forest masks, for forest management and protection
tasks on a regional level and in future also on a national level.
REFERENCES
Akca, D., 2005. Registration of point clouds using range and
intensity information. In: The International Workshop on
Recording, Modeling and Visualization of Cultural Heritage,
Ascona, Switzerland, May 22-27, in E. Baltsavias, A. Gruen, L.
Van Gool, M. Pateraki (eds.), Taylor & Francis/Balkema,
Leiden, pp. 115-126.
Artuso, R., St. Bovet, St., and Streilein, A., 2003. Practical
methods for the verification of country-wide terrain and surface
models. In: The Proceedings of the ISPRS working group III/3
workshop XXXIV-3/W13. 3-D reconstruction from airborne
laserscanner and InSAR data, 8-10 March, 2003, Dresden,
Germany.
Baltsavias, E., 1999. Airborne laser scanning: existing systems
and firms and other resources. ISPRS Journal of
Photogrammetry and Remote Sensing, 54, pp. 164-198.
Baltsavias, E., Gruen, A., Eisenbeiss, H., Zhang, L. and Waser,
L.T., 2008. High Quality Image Matching and Automated
Generation of 3D Tree Models. International Journal of Remote
Sensing, 29(5), pp. 1243 - 1259.
Brassel, P. and Lischke, H., 2001. Swiss National Forest
Inventory: methods and models of the second assessment.
Birmensdorf, Swiss Federal Research Institute WSL, 336 p.
Gruen, A. and Akca, D., 2005. Least squares 3D surface and
curve matching. ISPRS Journal of Photogrammetry and Remote
Sensing, 59(3), pp. 151-174.
Holmgren, J. and Persson, A., 2004. Identifying species of
individual trees using airborne laser scanner. Remote Sensing of
Emvironment, 90, pp. 415-423.
Hyyppä, J., Hyyppä, H., Inkinen, M., Engdahl, M., Linko, S.,
and Yi-Hong, Z., 2000, Accuracy comparison of various remote
sensing data sources in the retrieval of forest stand attributes.
Forest Ecology and Management, 128, pp. 109-120.
Küchler, M., Ecker, K., Feldmeyer-Christe, E., Graf, U.,
Küchler, H., and Waser, L.T. 2004. Combining remotely sensed
spectral data and digital surface models for fine-scale modeling
of mire ecosystems. Community Ecology, 5(1), pp. 55-68.
Lamonaca, A., Corona, P., Barbati, A. Exploring forest
structural complexity by multi-scale segmentation of VHR
imagery. Remote Sensing of Environment (2008),
doi:10.1016/j.rse.2008.01.017.
McCullagh, P. and Nelder, J.A., 1983. Generalized linear
models. Chapman and Hall, London, 511 p.
Naesset, E. and Gobakken, T., 2005. Estimating forest growth
using canopy metrics derived from airborne laser scanner data.
Remote Sensing of Environment, 96, pp. 453-465.
Scott, J.M., Heglund, P.J., Samson, F., Haufler, J., Morrison, M.,
and Wall, B., 2002. Predicted species occurrences: issues of
accuracy and scale, Island Press, Covelo, California, 868 p.
St-Onge, B., Jumelet, J., Cobello, M., and Vega, C. 2004.
Measuring individual tree height using a combination of
stereophotogrammetry and lidar. Canadian Journal of Forest
Research, 34(10), pp. 2122-2130.
Straub, B., 2003. Automated Extraction of Trees from Aerial
Images and Surface Models. In: The International Archives of
Photogrammetry and Remote Sensing, XXXIV-3/W8, pp. 157164.
1410
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
Waser, L.T., Küchler, M., Ecker, K., Schwarz, M., Ivits, E.,
Stofer, S. and Scheidegger, C. 2007. Prediction of Lichen
Diversity in an Unesco Biosphere Reserve - Correlation of high
Resolution Remote Sensing Data with Field Samples.
Environmental Modelling & Assessment, 12(4), pp. 315-328.
Waser, L.T., Baltsavias, E., Ecker, K., Eisenbeiss, H., Ginzler,
C., Küchler, M., Thee, P. and Zhang, L., 2008a. High-resolution
digital surface models (DSM) for modelling fractional
shrub/tree cover in a mire environment. International Journal of
Remote Sensing, 29(5), pp. 1261 - 1276.
Waser, L.T., Baltsavias, E., Ecker, K., Eisenbeiss, H.,
Feldmeyer-Christe, E., Ginzler, C., Küchler, M., Thee, P. and
Zhang, L., 2008b. Assessing changes of forest area and shrub
encroachment in a mire ecosystem using digital surface models
and CIR-aerial images. Remote Sensing of Environment, 112(5),
pp. 1956 - 1968.
Zhang, L. and Gruen, A., 2004. Automatic DSM generation
from linear array imagery data. In The International Archives of
Photogrammetry, Remote Sensing and Spatial Information
Sciences, XXXV, Part B3, 2004, pp. 128-133.
Zhang, L., 2005. Automatic Digital Surface Model (DSM)
Generation from Linear Array Images, PhD Thesis, Report No.
88, Institute of Geodesy and Photogrammetry, ETH Zurich,
Switzerland.
ACKNOWLEDGEMENTS
The study was carried out at the Swiss Federal Research
Institute WSL and funded by the Swiss Agency for the
Environment, Forest and Landscape (BAFU) and WSL. We
acknowledge Patrick Thee for his support in the field surveys.
Watt, P.J. and Donoghue, D.N.M., 2005. Measuring forest
structure with terrestrial laser scanning. International Journal of
Remote Sensing, 26(10), pp. 1437-1446.
1411
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B8. Beijing 2008
1412